README.md

Dataset Instruction

Directory Structure

root  
├── dataset 
│   ├── COCO     
│   │    │── captions_train2017.json    
│   │    │── captions_val2017.json
│   │    │── COCO_triplet_labels.npy    
│   │    └── images 
|   │         └── *.png        
│   ├── VG
│   │    │── image_data.json
│   │    │── VG-SGG-with-attri.h5
│   │    │── VG-SGG-dicts-with-attri.json
│   │    └── images
│   │         └── *.png 
│   ├── GQA
│   │    │── GQA_200_ID_Info.json
│   │    │── GQA_200_Train.json
│   │    │── GQA_200_Test.json
│   │    └── images
│   │         └── *.png
│   ├── CC
│   │    │── Train_GCC-training.tsv
│   │    │── cc_triplet_labels.npy
│   │    │── cc_meta_information.json
│   │    └── images
│   │         └── *.png  
│   ├── VG_Caption
│   │    └── region_descriptions.json

Training data (Caption dataset)

We use training datasets for COCO, CC, and Visual Genome caption datasets. You can conveniently download each set of training data using shell code. For detailed information, please refer to each following links.

# DATASET: COCO, CC, VG_Caption
bash dataset/{DATASET}/download.sh dataset/{DATASET}

Detailed Download instructions

• COCO: README.md
• CC: README.md
• Visual Genome Caption: README.md

Test data

For evaluation, we use Visual Genome (VG) and GQA datasets.

# DATASET: VG, GQA
bash dataset/{DATASET}/download.sh dataset/{DATASET}

Detailed Download instructions

• Visual Genome: README.md
• GQA: README.md